Wavelet analysis and multiscale pattern classification in wind engineering

Wavelet analysis was used to develop a multi-scale pattern detection and classification algorithm for intermittent time series, which are characterized by large transients separated by relatively quiescent periods. The detection algorithm is based on the dyadic wavelet transform, in which regions of sharp change in the signal are detected as having locally large wavelet coefficients that are contiguous across two or more dyadically-spaced wavelet scales at a given time index. Empirically-devised criteria are used to determine if the detected edge forms a boundary of a given transient or falls between other local edges, such that it is part of a longer transient. Traditional self-organizing pattern classification techniques, which cluster patterns into classes based on their similarity in feature space, were used to enable the detection of underlying structure in the observed transients. The resulting algorithm was tested successfully using time series consisting of artificial patterns and white noise. Then, roof-corner pressure measurements from Texas Tech University were processed to explore the applicability of multi-scale pattern detection and classification to field data. Finally, synthetic time series were constructed based on the patterns extracted from one of the pressure signals. These time series are expected to be useful as excitation inputs to computational structural models in parameter studies.